Why the study?
Hospital readmissions result in financial penalties from CMS, and predicting 30-day readmissions remains a difficult challenge in healthcare machine learning.
Can machine learning models using ICD9 embeddings, chart events, and demographics accurately predict 30-day hospital readmission in ICU patients?
Can machine learning models using ICD9 embeddings, chart events, and demographics accurately predict 30-day hospital readmission in ICU patients?
A Random Forest machine learning model utilizing ICD9 embeddings, chart events, and demographics achieved modest performance (AUC 0.66) in predicting 30-day hospital readmissions among ICU patients.
Modest ML performance should not yet guide ICU readmission prevention; leaves open whether refined embeddings or larger datasets can improve prediction.
Patient readmission to the hospital within 30 days or 365 days is a challenging problem for hospitals as they get penalized and in many cases the Center of Medicaid and Medicare (CMS) will not reimburse the hospitals for the costs associated with these readmissions. Although readmission prediction is a common problem in healthcare and has been addressed by the researchers in the machine learning community, it remains a hard problem to solve. The goal of the project proposed in this paper is to build a predictive model for 30-day readmission based on the Medical Information Mart for Intensive Care (MIMIC III) dataset, which contains admissions for intensive care unit (ICU) patients. We used ICD9 embedding's, chart events and demographics as features to train multiple classifiers including Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR) and Multi-Layer Perceptron (MLP). Best model, Random Forest, achieved 0.65 accuracy and 0.66 Area Under the Curve (AUC).
No takes yet. Share an insight, caveat, or question.
Assaf et al. (2020) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: